diagnostic decision support of a real-world satellite with lydia-ng
DESCRIPTION
Model-based diagnosis and AI-based automated diagnostic reasoning will be a full-success when we see this promising technology widely deployed everywhere---from mobile phones, to data centers and industry SCADA. In 2011 and 2012, an European start-up, General Diagnostics, in co-operation with leading European universities and a company with experience in PA of space systems, has won an ESA ITI project to make a step in real-world application of MBD to a real-world satellite. This presentation and the accompanying paper, published at AEROCONF'13, summarizes our experience from this project.TRANSCRIPT
Model-Based Diagnostic Decision-Support System for Satellites
Alexander Feldman, Helena Vicente de Castro,Arjan van Gemund, and Gregory ProvanMar 04, 2013 @ 09:45, in Lake/CanyonAEROCONF-2013, Big Sky, Montana
04/13/2023 2
Introduction
• History– LYDIA – diagnosis of Boolean circuits
• SAFARI – stochastic diagnosis• FRACTAL – disambiguation of diagnoses (reduce diagnostic uncertainty – entropy
based methods)• Beyond diagnosis – worst case sensor data – diagnosability (MIRANDA)
– A resistor network can be modeled with Boolean variables but that is very difficult, e.g., 4-bit multiplication tables
– LYDIA-NG – generalization to continuous variables• SAFARI-NG – greedy stochastic reasoning, but also other candidate generation policies• FRACTAL-NG – disambiguation
• The insight that allows generalization of Lydia to Lydia-NG is that simulation is a key-component in model-based diagnosis– What is the simulation step in the MBD circuit problem shown in the next
slide?
04/13/2023 3
GOCE
4
GOCE (cont.)
04/13/2023
04/13/2023 5
GENIUS Overview
04/13/2023 6
SCOS-2000 Integration
04/13/2023 7
GOCE Electrical Power System
04/13/2023 8
Fault Diagnosis
• Identify component(s) that are root cause of failure (error)
x, y: observation vectorsf: system function, fi: component functions
h: system health state vector, hi: component health variables
• Diagnose failure: solve inverse problem h = f-1(x,y)• Diagnosis: h2 = fault state, or h4 and h5 = fault state
y = f(x,h)xf1 f2
f4 f5
f3 healthy
faulty
04/13/2023 9
LYDIA-NG - Introduction
• LYDIA-NG is a diagnostic framework– Use cases:
• modeling of diagnostic systems (also an IDE)• running of diagnostic scenarios in batch mode (computation of
diagnostic metrics)• embedding in a SCADA system, e.g., a BMS
• Our view on the development of LYDIA-NG– collection of tools (simulators, translators, etc.)– simple interfaces (our users are not assumed to be expert
MBD users)– higher coding/documentation/knowledge dissemination
standards than typical projects
04/13/2023 10
LYDIA-NG Overview
• Core libraries– simulation
• SPICE• symbolic• ODEs/DAEs• Boolean circuits (old LYDIA heritage)
– diagnosis• forward reasoning (simulation for various health/fault states)• backward reasoning (residual analysis)
– disambiguation• entropy-based selection of tests• virtual sensors (by-product)
04/13/2023 11
State of LYDIA-NG Development
• LYDIA-NG has reached a milestone in diagnosing an electrical system– results show that the
software will be useful in practice
– releasing version 1.0 after fixing some bugs, documentation, and testing
– preview versions of the software continuously made available to the community for purpose of progress-tracking, collaboration, and planning
04/13/2023 12
LYDIA-NG Approach to Diagnosis
• The main idea is to run multiple-simulations simultaneously (each simulation reflects different health state)
• Choose those simulations that minimize some residual function
• Report diagnosis as a probability of each component being healthy/faulty
Model
Generate diagnostic
assumptions health1 health2 healthn
Simulate Simulate ...
Simulate
prediction1 prediction2 predictionn
Sensor
readings
Analyze residual
Analyze residual
... Analyze residual
candidate1 candidate2 candidaten
Choose plausible
candidates
set of plausible candidates
Compute probability for each component state
(healthy, faulty)
04/13/2023 13
GOCE Simulation
• Why SPICE– this is how electrical engineers simulate circuits
• very precise• fast• bug-free
• Caveats– going too detailed increases accuracy but this
takes computational time, it is too error-prone, time consuming (modeling effort), and is not needed for diagnosis (it can actually harm)
04/13/2023 14
Modified Nodal Analysis
04/13/2023 15
Fault Modeling
• Current sensor stuck-at OOB value• Voltage sensor stuck-at OOB value• Open-circuited heater loads• Switches stuck-at-opposite• Switches inverses intermittently• Degradation of Solar Panels
04/13/2023 16
LYDIA-NG Disambiguation
• Entropy-based methods for computing uncertainty of a component:
• Per system:
*
*
CxC
xCxCCU PrlgPr
COMPSC
CUCOMPS
U1
04/13/2023 17
Test Scenarios
ID Faults Classification Comment
TST1 No faults - Nominal Scenario
TST3One heater load (GCDE from HTRGRPA8) is open-circuited
Single Ambiguous fault -
TST4Continuation of TST3: Injection of Tests suggested by LNG
- Active testing to reduce candidates
TST7
Two heaters (GCDE & PXFA2_HPS from HTRGRPA8) open-circuited & two solar arrays (CBM-Mz & W-Mz) degraded
Abrupt Multiple Ambiguous
Simultaneously independent failures (highly unlikely case)
04/13/2023 18
Diagnoses and Metrics
• Evaluating the evaluator (comparing the performance of various diagnostic algorithms)
• Fault injection – deterministic assignment• Diagnosis
– was a set of deterministic diagnostic candidates– proposing a single probabilistic assignment
04/13/2023 19
Performance Metrics• Scenario detection accuracy, Mda
• Mfp (false positive scenario) is a scenario where diagnosis is produced but there is no injection
• Mfn (false negative scenario) is a scenario where diagnosis is not produced but there is an injection.
• Fault detection time, Mfd = FDT
• Fault Isolation time, Mfi = FDI
fpfnda MMM ,max1
04/13/2023 20
Performance Metrics (cont.)
• Classification errors
COMPSC
err xCxCIM )Pr()(
otherwise,0
)(,1)(
xCifxCI
04/13/2023 21
Performance Metrics (cont.)
• Classification errors• Uncertainty (entropy):
– the health state of each component is represented by a (random) variable
– entropy of random variable– entropy sum per system
Pr(R1 = 20) = 0.5,Pr(R1 = ∞) = 0,Pr(R1 = 1) = 0.5
Pr(R1 = 20) = 1/3,Pr(R1 = ∞) = 1/3,Pr(R1 = 1) = 1/3
Pr(R1 = 20) = 1,Pr(R1 = ∞) = 0,Pr(R1 = 1) = 0
04/13/2023 22
TST1: Nominal Scenario
– Nominal Scenario No faults injected• All heaters on demanded current of • All Solar arrays in MPPT
– Full luminosity• Main bus voltage equal to• Battery charging
04/13/2023 23
TST1: Nominal Scenario
04/13/2023 24
TST3: Single Ambiguous
– Start with nominal scenario No faults
– At some point the GCDE heater load is open-circuited
– In telemetry, value of the HTRGRPA8 LCL current sensor decreases of value
04/13/2023 25
TST3: Single Ambiguous
04/13/2023 26
TST3: Single Ambiguous
04/13/2023 27
TST3: Single Ambiguous
Fault injection:GCDE heater load OC
04/13/2023 28
TST4: Disambiguation of TST3
04/13/2023 29
TST4: Disambiguation of TST3
TST7: Multiple Ambiguous
Fault(s) injection
Command executed: GCDE_HTR Off
Command executed: IPCU_B_HTR2 Off
Command executed: EGGIF6_HTR Off
Command executed: TCVmZ_HTR Off
Command executed: TCVpZ_HTR Off
30
31
TST7: Multiple Ambiguous
Fault(s) injectionCommand executed: GCDE_HTR Off
Command executed: IPCU_B_HTR2 OffCommand executed: EGGIF6_HTR Off
Command executed: TCVmZ_HTR OffCommand executed: TCVpZ_HTR Off
Command executed: PXFA2_HPS HTR OffCommand executed: TCVpZ_HTR ON
Command executed: TCVmZ_HTR ONCommand executed: EGGIF6_HTR ONCommand executed: IPCU_B_HTR2 ON
Command executed: GCDE_HTR ON
32
GENIUS Results - I
Scenario Mfd[ms] Mfn Mfp Mda
TST1 − 0 0 1TST2 88 0 0 1TST3 88 0 0 1TST4 89 0 0 1TST5 90 0 0 1TST6 90 0 0 1TST7 93 0 0 1TST8 94 0 0 1TST9 84 0 0 1TST10 − − − −TST11 77 0 0 1
33
GENIUS Results - II
Scenario Mfi[s] Merr Mia Ment
TST1 − 0 289 0TST2 − 0 289 0TST3 505 1.45 287.55 0.004TST4 − 0 289 0TST5 − 1 288 0.0035TST6 101 1 288 0.0016TST7 505 3.96 285.04 0.0047TST8 − 1 288 0.0022TST9 − 1 288 0TST10 − − − −TST11 505 1.83 287.17 0.0025
04/13/2023 34
Conclusions
• A step toward applying MBD to a real-world artifact
• Our goals:– make diagnosis universal (to a large class of real-
world systems)– develop tools
• software• publications• public awareness of the importance of diagnosis
04/13/2023 35
Future (Present) Work
04/13/2023 36
AEROCONF-2013
• Read our paper:
A. Feldman, H. V. de Castro, A. J. C. van Gemund, & G. Provan,“Model-Based Diagnostic Decision-Support System for Satellites“, in Proc. IEEE Aerospace Conference 2013 - AEROCONF'2013,pp. 1-14, 2013
04/13/2023 37
Call for Participation
• Want to play with/test/develop LYDIA-NG?– LYDIA-NG is free for academic use/open-source– send an email to [email protected]
• Want to apply LYDIA-NG to your research/write a paper?
• Want to extend LYDIA-NG to solve #@! equations?
04/13/2023 38
Come to DX-2013
• Come to DX-2013, Oct 1-4, Dan Panorama Hotel, Jerusalem (http://dx-2013.org/)
• DXC-2013• synthetic track (ISCAS)• electrical system (ADAPT)• thermal fluid system (AHU-9)• software track